ADR-016: Ambiguity-Driven Architecture with Chain-of-Draft Integration

Status: Proposed Date: August 17, 2025 Decision Makers: PM, Chief Architect, Chief of Staff Classification: Revolutionary Bundle (includes Chain-of-Draft Integration, Ambiguity Assessment, Three-Tier Orchestration)

Context

Our research into agent architectures revealed a fundamental insight that challenges conventional wisdom: solution path ambiguity, not computational complexity, should drive architectural decisions for single vs. multi-agent systems.

This discovery emerged from:

  1. Analysis of Rahul Vir’s Framework 3: “Choose an Architecture Based on Problem’s Complexity” (which actually emphasizes ambiguity)
  2. Chain-of-Draft research showing 92% token reduction makes multi-agent systems economically viable
  3. Industry patterns from Anthropic, Microsoft, IBM, and LangChain

The convergence of these insights, combined with our MCP federation and spatial intelligence capabilities, creates a paradigm shift in how we architect Piper Morgan’s decision-making.

Economic Transformation Claims

The economic implications require verification per ADR-015:

Original Cost Optimized Cost Reduction Verification Status
Daily standups: $15 $0.75 95% Calculated from token reduction, needs production validation
Sprint planning: $125 $6.25 95% Calculated from token reduction, needs production validation
Quarterly analysis: $2,500 $125 95% Calculated from token reduction, needs production validation

Note: Cost reductions based on 92% token reduction from Chain-of-Draft paper (verified) applied to estimated token costs (unverified). Actual costs will vary based on model pricing and task complexity.

Decision

We will implement an Ambiguity-Driven Architecture that routes tasks based on solution path clarity rather than computational complexity, using Chain-of-Draft optimization to make multi-agent orchestration economically viable.

Core Components

1. Solution Path Assessment Framework

class AmbiguityAssessor:
    """Evaluates solution path clarity, not task complexity"""

    def assess_solution_clarity(self, task: Task) -> float:
        """
        Returns clarity score 0.0-1.0 based on:
        - Precedent existence (have we solved similar before?)
        - Step predictability (can we enumerate the steps?)
        - Success criteria clarity (do we know what "done" looks like?)
        - Domain consensus (is there an accepted approach?)
        """
        indicators = {
            'precedent_exists': self.check_precedent(task),
            'steps_enumerable': self.can_enumerate_steps(task),
            'success_measurable': self.has_clear_success_criteria(task),
            'domain_consensus': self.check_domain_consensus(task)
        }
        return self.calculate_clarity_score(indicators)

2. Three-Tier Orchestration Model

Tier 1: Clear Path (Score > 0.8)

Tier 2: Exploratory Path (Score 0.4-0.8)

Tier 3: No Clear Path (Score < 0.4)

3. Chain-of-Draft Integration

class DebateDrivenCoD:
    """Enables rapid multi-agent consensus at 1/20th token cost"""

    def compress_reasoning(self, agent_thought: str) -> str:
        """Compress to 5-word expressions per CoD methodology"""
        # Example:
        # Original: "I believe we should prioritize the authentication
        #           feature because security is critical for enterprise..."
        # Compressed: "authentication→security→enterprise→critical→priority"
        return self.extract_semantic_core(agent_thought, max_words=5)

    def orchestrate_debate(self, agents: List[Agent], problem: str):
        """Orchestrate compressed debate until consensus"""
        debate_rounds = []
        while not self.consensus_reached(debate_rounds):
            round_contributions = []
            for agent in agents:
                compressed = self.compress_reasoning(agent.reason(problem))
                round_contributions.append(compressed)
            debate_rounds.append(round_contributions)

        return self.synthesize_consensus(debate_rounds)

4. Economic Optimization

token_economics:
  traditional_debate:
    tokens: 50000
    cost: $2.50
    time: 45s

  debate_driven_cod:
    tokens: 2500  # 95% reduction [Verified: CoD paper]
    cost: $0.125  # 95% reduction [Calculated: Based on token reduction]
    time: 2.3s    # 95% reduction [Estimated: Needs measurement]
    quality_correlation: 0.96  # Only 4% quality loss [Source: CoD paper]

  confidence_note: "Token reduction verified in research, cost/time extrapolated"

Routing Examples

Task Clarity Score Route Rationale
“Fix typo in README” 0.95 Single agent Clear path, simple execution
“Generate test suite for auth module” 0.85 Single agent + CoD Clear path but complex
“Improve team morale” 0.3 Multi-agent debate No clear path, needs perspectives
“Optimize database performance” 0.7 Supervised multi-agent Partially clear, needs exploration
“Predict market disruption” 0.2 Full orchestration Highly ambiguous, emergent solution

Note: Clarity scores are illustrative. Actual scoring algorithm to be developed and calibrated through production usage.

Consequences

Positive

  1. Economic Viability: 92% token reduction makes multi-agent practical for routine tasks
  2. Better Decisions: Ambiguous problems get appropriate multi-perspective analysis
  3. Resource Efficiency: Clear problems avoid unnecessary orchestration overhead
  4. Adaptive Architecture: System matches approach to problem nature
  5. Paradigm Shift: Changes PM assistance from $2,500/quarter to $125/quarter

Negative

  1. Assessment Overhead: Must evaluate ambiguity before routing (adds ~100ms) [Confidence: Medium - Estimated based on similar classification tasks]
  2. Training Required: Team needs to understand ambiguity vs. complexity distinction [Confidence: High - Observed confusion in our own research process]
  3. Debugging Complexity: Multi-agent debates harder to trace than single agent [Confidence: High - Known challenge from industry experience]
  4. Quality Risk: 4% degradation from CoD compression [Confidence: High - Measured in CoD paper, but our implementation may vary]

Neutral

  1. Cultural Shift: Move from “how complex?” to “how clear is the path?”
  2. Metrics Evolution: New KPIs around ambiguity assessment accuracy
  3. Architecture Documentation: Must explain when/why multi-agent triggers

Implementation Plan

Phase 1: Foundation (Week 1)

Phase 2: CoD Integration (Week 2)

Phase 3: Routing Logic (Week 3)

Phase 4: Production (Week 4)

Alternatives Considered

Alternative 1: Complexity-Based Routing

Description: Use computational complexity as routing criterion Rejected Because: Our research shows clear complex tasks succeed with single agents

Alternative 2: Always Multi-Agent

Description: Use multi-agent for everything (Anthropic approach) Rejected Because: 15x token cost unnecessary for clear-path problems

Alternative 3: Always Single-Agent

Description: Force all problems through single agent (Cognition approach) Rejected Because: Genuinely ambiguous problems need multiple perspectives

References and Influences

Notes

This ADR represents a paradigm shift in our architectural thinking. The key insight—that solution path ambiguity rather than computational complexity should drive architecture—fundamentally changes how we approach PM assistance.

The economic implications are staggering: what was a $2,500 quarterly PMF analysis becomes $125, making sophisticated multi-agent analysis accessible for routine PM work.

Success metrics: